Beruflich Dokumente
Kultur Dokumente
Dept. of Geoinformation, Photogrammetry and Remote Sensing of Environment, AGH University of Science and
Technology, Krakow, Poland (galia, drzewiec, awrobel)@agh.edu.pl
ABSTRACT:
The main aim of the research described in the paper was analyse the accuracy of photointerpretation of impervious surface using
IKONOS images and its influence on the imperviousness factor determination. Two kinds of IKONOS image were chosen to tests:
panchromatic and colour pansharp. Airborne ortopho (pixel size of 0.2m) was applied as a reference. Six operators, digitised three
times, two kinds of IKONOS image, on the six test areas (300x300m). Accuracy analysis was performed applying different
parameters, among others: RMS and reproducibility (ISO 5725-2). Then, each test area was grided with 30m pixel size (simulation
Landsat image) and imperviousness factor was in each pixel determined. Mean error for PAN image was ca. 20% and for RGB
image ca. 10%.
1. INTRODUCTION and in the most cases such information is acquired from digital
aerial or satellite orthophotos. High resolution satellite images
In many cases traditional land-use / land cover map created are commonly used for this purpose. Here we comes to the
through classification of satellite images does not provide us question about the accuracy of these training and more
with information necessary for evaluation of changes occurring importantly validation (or control) datasets. In many cases high
in the landscape. The process of landscape urbanization can be resolution satellite orthoimages are used. Despite our efforts we
given as an example. In this case the changes may be twofold. werent able to find in literature any assessments of accuracy
What can change is not only the type of land-use (eg. from for imperviousness factor estimations based on photo
agriculture to residential area), but also the level of urbanisation interpretation of high resolution satellite imagery. Actually the
within the same land-use type. For former the traditional land- only information about the accuracy of photo interpretation
use map is enough, for latter may not. Because what changes based imperviousness data was find in Deguchi and Sugio
here it is not the land-use type, but the proportions of different (1994). They use the aerial photographs in different scales
kinds of land cover inside the same land-use class. Very (from 1:10000 to 1:23000) to obtain the reference dataset. They
detailed land-use / land cover map made from high resolution report the accuracy of the estimation of imperviousness factor
satellite images or air photos could be proposed in such a case, by visual interpretation of these photographs to be about 10 per
but such a map is very laborious and expensive when large area cent. We could expect similar or even worse accuracy of
is taken into consideration. Moreover, this kind of images may imperviousness factor derived by visual interpretation of high
be not available for past years. Continuum-based classification resolution satellite imagery. The verification of this assumption
of medium-resolution satellite images may be seen as a viable was set as a goal in research presented in our paper.
alternative (see e.g. Clapham 2003, Xian and Crane 2005, Xian
2006). As a result of such classification a map of
imperviousness factor is obtained. The imperviousness factor 2. ISO 5725-2 STANDARD AND ITS APPLICATION TO
can be defined as a percentage of the area (e.g. percentage of PHOTO INTERPRETATION
the image pixel) covered by impervious surfaces (such as roofs,
asphalt roads, parking lots, etc.). Acquisition of spatial data should be accompanied by
acquisition of information about their quality. In our opinion
Medium resolution satellite images have been used for the information about GIS data accuracy should be seen as one of
assessment of the ground surface imperviousness from 1970s the most important metadata, especially if the data are to be
(see Jackson 1975). Initially the methodology was based on used in financial context as penalty (e.g. Integrated
supervised or unsupervised image classification techniques, but Administration Control System IACS, in agriculture financial
because of the resolution of these images the results were often subsidies in EU) or taxation (e.g. cadastre, sewer waters). A
not satisfactory. Then many new approaches have been necessity for such information is also stressed in official
developed, including among others artificial neural network, regulations. In the Directive of European Council from 14
spectral mixture analysis or regression tree approach. A review March 2007 establishing the Infrastructure for Spatial
of up-to-date techniques can be found e.g. in Weng (2008). The Information in the European Community (INSPIRE) we can
accuracy of the imperviousness factor estimation reported in find the following statement: metadata in spatial database shall
different studies is usually better then 20 per cent. include information on the quality and validity of spatial data
sets (Chapter II, Metadata, Article 5, p.2 c). GIS data metadata
Regardless the approach applied, the information about the as defined in ISO 19113 standard contain among others: quality,
impervious surfaces acquired in the field or from higher spatial accuracy, temporal accuracy and thematic accuracy.
resolution data is needed as a training (or calibration) data and
also for accuracy assessment. The field data are rarely available
1329
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
1330
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
measurements and singular measurement. Instead of mean, the absolute RMS [sq m], relative area error (RRMS Relative
true value should be used, if it is accessible. Alternative method, Root Mean Square = RMS/reference area with value: 0-100%
to RMS, could be also involved to the accuracy assessment, or 0-1) was analysed.
especially if the experiment is properly prepared and
measurements are repeated by a few operators. In this case, the 4.2 Part 2 - influence of the photointerpretation accuracy
experiment is similar to laboratory measurements usually made on the imperviousness factor calculated in simulated
in chemical research, precisely determined by ISO standard. Landsat pixel
According ISO 5725 Accuracy (trueness and precision) of
measurement methods and results, among others, the following Landsat classification (e.g. unmixing) requires training area,
notation is used: sample, operator, laboratory, equipment, obtained from field surveying or from photointerpretation of
repeatability and reproducibility. As a sample, in our research VHR images (e.g. IKONOS). In our research, we ask the
we understand digitised impervious surface on IKONOS, on question: how does the photointerpretation accuracy influence
each of test area. In the experiment, 6 operators have been on the percent of impervious surface in Landsat pixel?
taking part using 2 equipments (IKONOS PAN and RGB). Each
operator is understood as a laboratory, so we have 6 For each test area, grid of 30m cell size was simulated. Then the
laboratories. Each test area has, as a reference, the impervious impervious surface area was calculated in the grid. The
surface area, digitised on the aerial ortophoto. Six operators procedure was performed for all measurements (18 observations
digitised PAN/RGB 3 times, so we have 18 measurements for for each test area) and for the reference. In each simulated
each test area. Finally, 108 measurements were analysed Landsat pixel, area of impervious surface and percent of
because of 6 test area and 18 measurements of one test area. At impervious surface area in the pixel was calculated for each
the beginning, outliers were found for PAN and RGB images. measurement and for the reference. The percentage of
According ISO 5725 outliers are found using graphical or impervious surface area in the pixel is called: imperviousness
numeric method, Corchan/Grubbs tests. In our research group, factor. Reference area and measured area were analysed in 100
there was no statistician so we identified outliers in traditional pixels for the test area. Analysis was made in two aspects:
way, as used usually in surveying. Relative differences between comparison to the relative area error (RRMS) calculated
reference area and each measured area were calculated and generally for all test areas (data processing described in 4.1 Part
histogram of the error was prepared for PAN and RGB (Figure 1) and evaluation of the absolute and relative error of
4). Outliers were defined as 5% of the external measurements imperviousness factor. Firstly, RRMS of the impervious area
on the histogram. If at least one of measurement is marked as was calculated for each pixel. Then the differences between
an outlier, all group of measurements (3 repetitions) are percent of impervious cover of the pixel basing on observations
discarded. After outliers removing, accuracy analysis was and on reference were calculated. Then, absolute RMS of
performed. imperviousness factor was calculated for each pixel. Finally,
RMMS of impervious surface area was calculated for all pixels.
Assuming y, as a result of measurement of the impervious Accuracy analysis was performed on the Landsat pixel level in
surface area for each test area for each measurement, the 10 groups of imperviousness factor varies from: from 0% to
following can be written: 100% by each 10%.
y = m + B+ e 0 0.41 0.03 0
where:
m expected value (reference impervious area),
B bias in repeatability condition (difference between measured
area and reference area),
e - random error in repeatability condition. .05 0.51 0.03 0
Variance of B describes between laboratories variance:
var (B) = 2group
where:
group standard deviation (between laboratories).
Variance in repeatability condition for one laboratory is defined .37 0.16 0.4 0.79
as following:
var (e) = 2l
l standard deviation within group (laboratory), calculated for
the test area digitized 3 times by one operator.
Each operator (laboratory) is described by standard deviation l. .15 0 0.38 0.65
For all operators average variance is calculated, called variance
of repeatability:
Figure 3 Part of test area no. 1 with overlaid all digitised on
2repet=var(e) = 2l .
Pansharp colour IKONOS area; each pixel is labelled by
Finally, accuracy is defined as standard deviation of
percent of impervious surface [0,1].
reproducibility and it is the sum of the between groups variance
and the within groups variance:
2reprod = 2group + 2repet
5. RESULTS
where:
group- standard deviation in group (one for one test area), Initially, all observations for PAN and RGB were statistical
repet- average standard deviation of repetibility (average of six analysed. Histogram of relative area error for PAN and RGB is
standard deviations for each operator). on the figure (Figure 4) presented. Mean of the difference
Results of photointerpretation were validated using ISO between observed area and reference area (bias of the method)
standard and in traditional way basing on RMS. Besides
1331
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
calculated for all observations receive for PAN: -0.11 and for Test Refere reprod
RGB: 0.02. Standard deviation is respectively: 0.23 and 0.17. area nce (mean obs.- (mean.) RRMS
----------
no area ref. area) [%] ref. area (ref.
0.60 [m2] [%] [%] area)
0.50 [%]
1 23103 0.126 0.118 0.105 0.180
fequency [%]
0.40
0.30
PAN 2 14805 -0.026 0.184 0.114 0.168
0.20
RGB
3 52925 -0.025 0.069 0.060 0.070
4 51194 -0.001 0.058 0.056 0.054
0.10
5 19350 -0.051 0.117 0.097 0.113
0.00
6 25664 0.151 0.102 0.105 0.180
(-0.1;0.1]
(-0.6;-0.5]
(-0.5;-0.4]
(-0.4;-0.3]
(-0.3;-0.2]
(-0.2;-0.1]
(0.1;0.2]
(0.2;0.3]
(0.3;0.4]
(0.4;0.5]
(0.5;0.6]
(0.6;0.7]
mean. 0.029 0.108 0.089 0.127
relative error [%] Table 4. Accuracy analysis RGB bias () , standard deviation
(), reproducibility (reprod) and RRMS
Figure 4 Relation between relative area error [%] and its
frequency (histogram of errors for all data). PAN
0.30
Table 2. Accuracy analysis RGB repeatability, reproducibility Figure 6 Relation between relative area error, RRMS [%] and
and RRMS test area [ sq m], RGB
Table 3. Accuracy analysis PAN bias () , standard deviation Figure 7 Relation between relative area error RRMS [%] and
(), reproducibility (reprod) and RRMS area
[sq m], PAN, RGB (in simulated Landsat pixels, 30m)
1332
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
Observations were analysed to identify and discard outliers: Before statistical analysis, values of RMS for all test area and
- PAN: 8 outliers (4 groups 12 measurements), all pixels, were combined into 10 groups depending of the mean
- RGB: 5 outliers (3 groups 9 measurements) imperviousness factor from 0 to 100% by each 10%. Bias, and
RMS of imperviousness factor are presented in 10 groups for
After outliers discarding standard deviations received 0.20 PAN (Figure 8) and RGB (Figure 9) observations. Maximum
(PAN) and 0.14 (RGB). Repeatability, reproducibility and RMS bias in PAN measurements was slightly more then -30%, it
for the test areas are for PAN in table (Table 1) presented and in means that operators underestimated impervious areas.
table (Table 2) for RGB. Comparison between: bias (average Absolute RMS increases with increasing of imperviousness
observed area minus reference area), standard deviation (from factor even more then 35% on IKONOS PAN. IKONOS
the average observed area), reproducibility and RMS is possible PANSHARP (RGB) allowed obtaining much better results
in tables (Table 3, Table 4). Relationships between (Figure 9), small, neglected bias and RMS in some cases only
reproducibility, RRMS and the test area are shown on the slightly more then 10%.
diagrams (Figure 5 and Figure 6). Mean value of reproducibility
for PAN is equal: 13.3% and respectively RRMS: 21.6%. Better Finally, the RRMS of impervious surface area was calculated in
results were obtained for RGB, mean reproducibility is equal: each pixels for comparison to the RRMS of impervious surface
8.9% and respectively RRMS: 12.7%. Photointerpretation of area calculated for all test area (compare Figure 5, 6 and Figure
PAN was biased more than RGB (see Table 3 and Table 4). 10,11). In the analysis the value of the impervious area should
Mean bias on PAN for test areas 1,3,4,6 (digitized impervious be take into consideration, Fig, 5, 6: 10 000 60 000 sq m, and
surface) was -18% and for test areas 2, 5 (digitized pervious Figure 10, 11: 0 900 sq m.
surface): +12.3% (average error of in Table 3 is 16%). It
means that in photointerpretation on PAN impervious surfaces
0.40
were underestimated in compare to the reference. This
relationships is not observed on the RGB, bias is smaller than 0.30
on PAN, varies from minus to plus values and mean of RRMS 0.20
is 2.9% (Table 4). Relationships between the RRM and test area 0.10
are presented for PAN and RGB on the diagrams, respectively BIAS PAN
0.00
on Figure 5 and Figure 6. The bias, appearing in the PAN RMS PAN
observations, is possible to be seen on the figure Figure 5 as a
)
)
-0.10
.1
.2
.3
.4
.5
.6
.7
.8
.9
.0
;0
;0
;0
;0
;0
;0
;0
;0
;0
;1
shift of RRMS up to the reproducibility (ISO). This
[0
.1
.2
.3
.4
.5
.6
.7
.8
.9
[0
[0
[0
[0
[0
[0
[0
[0
[0
-0.20
phenomenon is not, in this scale, observed in RGB observations
-0.30
(compare Figure 5 and Figure 6). For example, RRMS for the
test area 3 and 4 is almost equal to the reproducibility, and the -0.40
bias is small.
The figures: Figure 7 - Figure 11 show the accuracy analysis of Figure 8 Relation between the absolute errors of
data, in simulated Landsat pixels (30m). Relation between the imperviousness factor: RMS, bias and the imperviousness factor
RRMS, relative area error of impervious surface in the pixel, combined into 10 groups, PAN
and reference area of impervious surface is presented on
diagram (Figure 7). Usually, the absolute area error increases 0.40
with increasing of the measured area, but simultaneously
0.30
RRMS is decreasing. The relationship is valid for the area from
digitalisation of remote sensing images or surveying. It is easy 0.20
-0.10
the tendency of decreasing of RRMS with the area is slightly to
.1
.2
.3
.4
.5
.6
.7
.8
.9
.0
;0
;0
;0
;0
;0
;0
;0
;0
;0
;1
[0
.1
.2
.3
.4
.5
.6
.7
.8
.9
[0
[0
[0
[0
[0
[0
[0
[0
-0.20
observations (Figure 6). The total measured impervious areas in -0.30
test areas are varying between 10 000 60 000 sq m, and the
-0.40
RRMS varies from 6%, in RGB, to even 25 % in PAN
observations.
Figure 9 Relation between the absolute errors of
Analysing the same relationship in grid of 30m, we could obtain imperviousness factor: RMS, bias and the imperviousness factor
even huge error for small areas heading to zero, because in this combined into 10 groups, RGB
case RRMS heads to infinity. Maximum area in grid 30m is
9000 sq m, so in Landsat pixel impervious area varies from 0
to 900m, and RRMS varies in range of zero to infinity. But the
decreasing tendency with area is noticeable on Figure 7 , even
in the cloud of points. On the other hand we use results of
photointerpretation as a test area in Landsat image classification.
Therefore, we compared imperviousness factor calculated for
pixel on the basis on all observations and on the reference.
Usually we assume result of one photointerpretation as a
reference. In our research we have 18 measurements because 6
operators digitised the area 3 times.
1333
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
)
.1
.2
.3
.4
.5
.6
.7
.8
.9
.0
;0
;0
;0
;0
;0
;0
;0
;0
;0
;1
panchromatic images were significant worse then obtained from
[0
.1
.2
.3
.4
.5
.6
.7
.8
.9
[0
[0
[0
[0
[0
[0
[0
[0
[0
the colour images. Some explanation might be the spectral
range of panchromatic channel of IKONOS covering also infra
Figure 10 Relation between RRMS and standard deviation of red, vegetations on the PAN image is bright and might be
RRMS of impervious surface area and the imperviousness misrecognised as a bright concrete cover (or against). Generally,
factor combined into 10 groups, PAN colour image contains more information useful for impervious
surface recognition.
1
0.9 ACKNOWLEDGMENTS
0.8
0.7 This research was done in frame of the project Multitemporal
0.6
RRMS remote sensing imagery based evaluation of spatial changes of
0.5
0.4
SD (RRMS) land-use and landscape functions for landscape planning
0.3 activities support; (N526029 32/2621) financed by the Polish
0.2 Ministry of Science and Higher Education.
0.1
0
)
REFERENCES
.1
.2
.3
.4
.5
.6
.7
.8
.9
.0
;0
;0
;0
;0
;0
;0
;0
;0
;0
;1
[0
.1
.2
.3
.4
.5
.6
.7
.8
.9
[0
[0
[0
[0
[0
[0
[0
[0
[0
1334